Emergency networking method fusing pre-disaster priori and boundary condition graph neural network

Abstract

This invention discloses an emergency networking method integrating pre-disaster prior knowledge and edge-conditional graph neural networks, including step S1, pre-disaster preparation: pre-disaster state quantification and spatiotemporal redistribution of computing power; step S2, post-disaster instantaneous survival verification and prior activation; step S3, post-disaster perception: high-dimensional topological state perception based on edge-conditional graph neural network EC-GNN; and step S4, post-disaster decision-making: dynamic evolution of strategy based on counterfactual differential reward mechanism. This invention, through the prior solidification of daily states, allows distributed nodes to directly enter a game state with predispositions during a disaster, eliminating the long and extremely power-consuming random trial-and-error period. The innovative EC-GNN edge feature modulation mechanism realistically reproduces the erosion of information entropy caused by spatial fading, avoiding the situation where all agents output homogeneous erroneous instructions in congestion scenarios. Through the counterfactual differential reward mechanism, the improvement of local congestion by a single node is accurately evaluated.

Description

Technical Field

[0001] This invention belongs to the field of mobile communications, specifically relating to an emergency networking method that integrates pre-disaster prior knowledge and edge conditional graph neural networks for mobile terminals with D2D (device-to-device) communication capabilities, such as smartphones, wearable devices, drones, vehicle-mounted communication nodes, and customized rescue terminals. Background Technology

[0002] With the development of mobile ad hoc networks and drone technology, building emergency communication networks using mobile terminals has become an important means of disaster relief. However, in extreme scenarios where sudden natural disasters cause infrastructure damage, existing technical solutions still have shortcomings.

[0003] First, existing technologies tend to use edge conditional graph neural networks to extract spatial correlations and characterize channel details. However, in distributed implementations, the need to frequently exchange high-dimensional hidden layer states leads to a huge message passing overhead [SHEN Y, ZHANG J, SONG SH, et al. Graph Neural Networks for Wireless Communications: From Theory to Practice [J]. IEEE Transactions on Wireless Communications, 2023, 22(5): 3554-69.].

[0004] Secondly, when performing collaborative tasks, terminals in disaster areas must strike a balance between energy efficiency and performance. Existing MARL resource scheduling schemes often focus on throughput or latency optimization, neglecting the hardware lock-in risk caused by frequent switching of heterogeneous protocol stacks and the "node sudden death" caused by rapid battery depletion [W. Farag, "Multi-Agent Reinforcement Learning using the Deep Distributed Distributional Deterministic Policy Gradients Algorithm," 2020 International Conference on Innovation and Intelligence for Informatics, Computing and Technologies (3ICT), Sakheer, Bahrain, 2020, pp. 1-6]. Existing technologies lack a decision-making mechanism that can explicitly consider the hardware survival layer, communication performance layer, and protocol stack stability, and achieve multi-objective Pareto trade-offs.

[0005] Finally, in multi-agent collaborative learning, although the counterfactual credit allocation mechanism proposed by existing technical solutions lays a theoretical foundation for distributed decision-making [J. Li, C. Shen and K. Zhang, "Convolutional neuralnetwork-based channel allocation and resource optimization for D2D communications," 2022 6th International Conference on Wireless Communications and Applications (ICWCAPP), Haikou, China, 2022, pp. 109-113], in the highly dynamic network environment after a disaster, due to severe time delay jitter, the traditional static baseline is prone to causing the reward signal signal-to-noise ratio to be too low.

[0006] Disadvantages of existing technology: Feature oversmoothing and physical distortion: Existing GNN operators directly perform linear superposition of neighbor features. At locations with high node density, the agent's hidden representations quickly converge, losing its ability to make differentiated decisions. Furthermore, they ignore the distorting effect of physical electromagnetic fading on the message passing process between graph nodes.

[0007] The cost of a "cold start" after a disaster is unbearable: Existing solutions require the network to train the model by randomly exploring the unknown environment after a disaster, and the trial and error in the early stage will cause a large number of nodes to waste their power.

[0008] It is highly susceptible to falling into a zero-sum game: Existing centralized or equal-distribution reward mechanisms cannot effectively punish selfish high-power transmission behavior, nor can they fundamentally curb MAC layer broadcast storms under high density.

[0009] In summary, existing technologies face the following core challenges when addressing the issue of building mobile ad hoc networks (MANETs) locally using a massive number of existing mobile terminals after the destruction of disaster infrastructure (base stations, etc.): (1) Severe conflict between node energy consumption and topology connectivity: The power supply of terminals in the disaster area is extremely limited. If the maximum transmission power is used to ensure connectivity, it will lead to network cascading power outages; if the power is reduced, it will lead to topology fragmentation.

[0010] (2) MAC layer broadcast storm in high-density scenarios: In densely populated disaster-affected areas, the disorderly channel competition of massive nodes will frequently trigger CSMA / CA exponential backoff, causing the collision rate to surge and the effective throughput of the system to approach zero.

[0011] (3) The “cold start” and perception limitations of distributed algorithms: In the traditional MARL algorithm, distributed nodes face partial observability (can only perceive single-hop neighbors) and are prone to getting trapped in local suboptimal conditions; while the introduction of traditional graph neural networks (GNN) can extract features, it is very easy to “over-smoothing” in dense disaster networks. Moreover, traditional algorithms start from scratch after a disaster, resulting in high long convergence delays, which cannot meet the timeliness requirements of emergency communication.

[0012] In view of the above-mentioned shortcomings of the existing technology, the present invention is proposed. Summary of the Invention

[0013] The purpose of this invention is to provide a MARL emergency communication optimization method that integrates pre-disaster state priors with edge-conditional graph neural network enhancement. It avoids cold starts by redistributing computing power before a disaster, overcomes spatial awareness barriers and feature oversmoothing by using EC-GNN, and encourages the emergence of implicit spatial reuse strategies in the distributed network through a counterfactual reward mechanism, thereby maximizing network lifetime and data delivery rate.

[0014] Beneficial effects: Compared with existing MARL topology control and GCN converged networks, this invention has the following significant advantages: Completely break through the "cold start" bottleneck and achieve "plug and play" after disaster: By solidifying the prior state of daily life (by advancing the process of finding computing power centers to before the disaster), distributed nodes can directly enter the game state with their own biases at the moment of disaster, eliminating the long and extremely power-consuming random trial and error period.

[0015] Effectively curb the "feature oversmoothing" disaster in high-density networking: The innovative EC-GNN edge feature modulation mechanism realistically reproduces the erosion of information entropy caused by spatial fading, avoiding the occurrence of homogeneous erroneous instructions output by all agents in congested scenarios, and greatly improving the spatial perception accuracy of the model in large-scale networks.

[0016] Achieving distributed "implicit spatial multiplexing" to eradicate broadcast storms: Through a counterfactual differential reward mechanism, the improvement of a single node on local congestion is accurately evaluated. In highly congested scenarios, this mechanism can prompt a large number of non-backbone nodes to actively choose a low-power silent strategy, spontaneously relinquishing valuable CSMA / CA air channel resources to key backbone nodes, achieving an effect similar to frequency division / space division multiplexing without centralized control.

[0017] Terminology Explanation: MANETs, ​​Mobile Ad-hoc Networks, are temporary, autonomous systems consisting of a group of mobile nodes equipped with wireless transceivers, characterized by a decentralized architecture and multiple hops.

[0018] MARL, Multi-Agent Reinforcement Learning, studies the algorithmic architecture of how multiple agents learn optimal policies through trial and error and interaction in a shared environment.

[0019] EC-GNN, Edge-Conditioned Graph Neural Network.

[0020] Dec-POMDP, or Decentralized Partially Observable Markov Decision Process, is a framework for mathematically modeling the game processes of multiple distributed agents that lack a central authority and a global state view.

[0021] CSMA / CA, Carrier Sense Multiple Access / Collision Avoidance, is a media access control protocol widely used in wireless local area networks. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 This is a flowchart of the emergency networking method of the present invention that integrates pre-disaster prior knowledge and edge condition graph neural network; Figure 2 This diagram illustrates the mapping relationship and message passing between the EC-GNN algorithm's abstract spatial feature aggregation and the physical terminal's disaster area environment. Figure 3 This is a diagram of the cross-temporal and spatial strategy execution and post-disaster purely distributed online fine-tuning architecture of the "cloud-based holographic pre-simulation - edge-based real-world evolution" of this invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0024] Figure 1 This is a flowchart of the emergency networking method of the present invention that integrates pre-disaster prior knowledge and edge condition graph neural network; Figure 2 This is a schematic diagram illustrating the mapping relationship and message transmission between the abstract spatial feature aggregation of the EC-GNN algorithm of this invention and the disaster area environment of the physical terminal; Figure 3 This is a diagram of the cross-temporal and spatial strategy execution and post-disaster purely distributed online fine-tuning architecture of the "cloud-based holographic pre-simulation - edge-based real-world evolution" of this invention, combined with... Figure 1 , Figure 2 as well as Figure 3 The present invention is described in detail below: This invention is divided into two coherent phases on a timeline: "pre-disaster prior modeling" and "post-disaster dynamic decision-making." At the algorithmic level, it combines Multi-Agent Reinforcement Learning (MARL) with Edge Conditional Graph Neural Network (EC-GNN) that deeply integrates physical characteristics. It proposes a mobile terminal emergency networking method based on pre-disaster prior modeling and EC-GNN enhancement, transforming the complex post-disaster channel fading and congestion penalties into a decentralized partially observable Markov decision process (Dec-POMDP).

[0025] The specific implementation method is as follows: Step S1, Pre-disaster preparation: Quantification of pre-disaster status and spatiotemporal redistribution of computing power; To avoid the system having to start from scratch in an unknown environment after a disaster (leading to high convergence delays and power depletion), this invention proposes to make preparations in advance during normal operation, specifically including: Step S10, Quantification of multi-dimensional physical indicators: During daily operation, continuously collect six core physical indicators of the terminal: absolute battery capacity, computing power, transmission power, connection stability, throughput, and packet loss rate. Multidimensional physical index extraction: During daily operation, the system collects the core index vectors from the terminal. Corresponding node Absolute battery capacity Computational ability Transmission power Connection stability Throughput and packet loss rate And perform extreme value normalization processing and map to Interval.

[0026] Step S11, Prior weight solidification: Calculate the prior score of the terminal's comprehensive performance using the Analytic Hierarchy Process (AHP), and generate an initial feature interaction prior weight matrix accordingly. Solidify the communication potential into a prior tensor and distribute computing power to edge nodes in advance. AHP Consistency Assessment and Prior Score Generation: Constructing a Positive-Negative Judgment Matrix By solving for the satisfaction The largest eigenvalue and its corresponding weight vector, satisfying the consistency ratio Under the premise of [the above], the prior score of the overall performance of the computing terminal In the formula Represents a positive-negative judgment matrix; This represents the weight vector corresponding to the indicator. This represents the largest eigenvalue of the judgment matrix; Represents the calculated nodes The comprehensive performance prior score.

[0027] Topological prior weight solidification: Based on the prior scores of the overall performance of the nodes, the initial feature interaction weight matrices used to guide the perception layer are calculated. And the action bias weight vector used to guide the decision-making level For any pair of nodes (i,j) in the network topology that may generate data interaction, their initial feature interaction weights are... It can be defined as

[0028] In the formula, Represents a node For nodes Initial feature interaction weights; and This is a constant adjustment coefficient; and They are nodes and nodes The overall performance prior score; and They are nodes and nodes The multidimensional physical feature vector; This refers to the regional correlation coefficient. Represents a node The set of single-hop direct neighbor nodes; This is the index variable for traversing the neighbor set. This function is not only proportional to the product of the prior scores of the two endpoints, but also integrates the cosine similarity of multidimensional spatial features and the regional correlation coefficient. By solidifying the communication potential into a prior tensor, it distributes computing power across the spatial axis to the edge nodes, thus solving the cold start problem of the model after a disaster.

[0029] It also includes step S12, cloud-based offline pre-training; In the pre-disaster phase, intensive pre-training is conducted. Relying on cloud computing power, a massive number of simulated disaster scenarios are constructed using domain randomization technology. The Actor network and Critic network complete global convergence of centralized training and distributed execution in the cloud, resulting in a mature basic policy model, which is pre-installed on all mobile terminals. Step S2, post-disaster instantaneous survival verification and prior activation, i.e., the heartbeat probe mechanism.

[0030] In the moments after the disaster: heartbeat probes and rapid topology trimming; The network is highly susceptible to fluctuations when switching from normal to post-disaster status. This invention designs a negative migration prevention mechanism, namely: Minimalist Heartbeat Verification: At the moment of a disaster, the system will not blindly activate the pre-disaster topology directly, but will trigger a low-overhead "millisecond-level survival probe". Each node will broadcast a minimal (e.g., 1 byte) heartbeat packet.

[0031] Avoiding convergence oscillations: If a node does not receive a heartbeat response from a high-potential node before the disaster, the system will instantly reduce its prior confidence to zero at the physical level and quickly remove the damaged node to prevent convergence oscillations in the early stages of the network.

[0032] At the moment the disaster occurred ( The system does not blindly and directly use the pre-fixed prior weight matrix. Instead, it triggers a very low-overhead "millisecond-level survival probe." Specifically, each node broadcasts a minimal heartbeat packet (such as a 1-byte pulse signal) carrying its unique identifier (MAC / ID) on the control channel. If a node... No high-potential nodes were received before the disaster within the set time window. If the heartbeat receipt is received, its instantaneous liveness indicator function will be forcibly reset to zero (i.e., modified). This allows for the instant removal of damaged nodes at the physical level, preventing initial convergence oscillations in the network.

[0033] Step S3, Post-disaster perception: High-dimensional topological state perception based on edge conditional graph neural network (EC-GNN); Ordinary graph neural networks (GNNs) are prone to "feature oversmoothing" in densely populated disaster areas, causing the agent to lose its differentiated decision-making ability. To address this, this invention designs a physically reconstructed edge conditional graph neural network, EC-GNN. By integrating electromagnetic physics features, EC-GNN uses the "instantaneous signal-to-noise ratio" and "end-to-end queuing delay" of the physical layer as side conditions, and utilizes a multilayer perceptron (MLP) for feature cascading and modulation. This physical reconstruction realistically reproduces the erosion of information entropy caused by electromagnetic fading.

[0034] Extremely low-overhead quantization interaction: To prevent high-dimensional feature interaction from exhausting the bandwidth of the disaster area, message broadcasting is designed as an "event-triggered" mechanism (triggered only when there are substantial changes such as drastic fluctuations in the link signal-to-noise ratio), and high-dimensional floating-point numbers are compressed into low-bit (such as 4-bit or 8-bit) discrete representations before being sent.

[0035] Spatiotemporal smooth transition: A control factor that decays exponentially with time step is constructed, and the topological prior extracted before the disaster is dynamically fused with the real-time channel state after the disaster to achieve a smooth transition from prior-dominated to real-time dynamic-dominated.

[0036] Specifically: After the disaster, each mobile terminal (smart agent or node) Under limited observation conditions, EC-GNN is used to extract global high-dimensional space belief states.

[0037] Step S30, local observation vector construction; intelligent agent Sniffing to obtain observation vectors Self-characteristics Includes normalized power, protocol status, node degree; edge features Includes instantaneous signal-to-noise ratio of physical links End-to-end queuing delay .

[0038] Step S31, based on event triggering and low-bit quantization of side conditional message interaction: To avoid exhausting the precious physical channel bandwidth of disaster areas through the interaction of high-dimensional feature vectors in GNNs, this invention introduces a communication overhead penalty and dimensionality reduction mechanism. Specifically, in terms of feature dimensionality reduction, a multilayer perceptron is constructed to map the concatenated neighbor node state vectors and link feature vectors into a new message carrier to achieve dimensionality reduction. Regarding communication overhead penalty, a penalty term for performing high-energy-consuming actions or locking heterogeneous protocol stacks is explicitly introduced into the counterfactual differential reward mechanism. This aims to curb ineffective feature broadcasting at the level of reinforcement learning objectives.

[0039] The event triggering is as follows: the agent only triggers physical layer broadcast interaction of state information when there is a substantial change in the local network topology, such as when the link signal-to-noise ratio fluctuates drastically beyond the threshold, or when there is a sudden change in the degree of a single hop.

[0040] The aforementioned low-bit quantization: Before broadcasting, the transmitting end uses a local quantizer to convert the high-dimensional floating-point features... The signaling overhead is compressed into a low-bit (e.g., 4-bit or 8-bit) discrete representation. After receiving the signal, the receiver performs dequantization and then generates a side-conditional modulation message through a multilayer perceptron. This reduces the control signaling overhead by several orders of magnitude.

[0041]

[0042] In the formula, This represents the index of the aggregation layer number in an edge-conditional graph neural network (EC-GNN). Indicates the first The edge-conditional modulated message feature vector generated by the layer; This represents the mapping function used to process message vectors in a multilayer perceptron network. This represents the dequantization operation function; Representing neighboring nodes In the The high-dimensional hidden state belief vector output by the layer.

[0043] Step S32, Dynamic attention fusion and prior confidence time decay: To address the negative migration caused by backbone node damage, a time-stepping mechanism is introduced. Control factor of exponential decay The probe liveness indicator function. Control factors The real-time attention score is weighted and fused to generate a dual-mode dynamic attention coefficient:

[0044] In the formula, Indicates the first Layer nodes For neighboring nodes The dual-mode dynamic attention coefficient; and These represent the unnormalized real-time attention scores calculated from the local channel state; This represents the initial maximum control factor constant; This represents the decay coefficient, indicating the rate at which the control factor decreases with time step. Represents a node The set of single-hop direct neighbor nodes; To traverse the set of neighbor nodes index variables; and This is the heartbeat probe survival indicator function; its value is 1 if the node is alive, and 0 otherwise. The latter part of the formula introduces the survival indicator function to... By performing masking operations and renormalization, the prior weights are accurately redistributed proportionally among the surviving nodes.

[0045] Step S33, Self-Embedding Reinforcement and Residual Normalization: In disaster areas where energy replenishment is difficult in a timely manner, the agent's perception of its remaining power level cannot be diluted by massive amounts of neighbor interference messages. Therefore, the system provides the agent with an independent, fully connected channel to strengthen its perception of its remaining power level. Specifically, the agent first constructs a system containing normalized remaining power levels. Current status of heterogeneous communication protocols and local node degree features Its own feature vector:

[0046] This inherent feature is utilized by a dedicated weight matrix when passing through the network. Independent linear mappings are performed, residual connections are introduced, and LayerNorm layer normalization operations are introduced to stabilize gradient backpropagation in deep networks. The specific iterative update rule of the layer is described as follows:

[0047] In the formula, This represents the mapped output of an independent, fully connected channel. (The process has been described.) The iterative aggregation of layered EC-GNNs rigorously up-maps low-dimensional local observations that were originally limited by single hops to include global dimensions. High-dimensional belief state vector influenced by topological jumps ...

[0048] Step S4, Post-Disaster Decision Making: Dynamic Evolution of Strategies Based on Counterfactual Differential Reward Mechanism

[0049] Step S40, Action Mapping: The agent uses the Actor network to Mapped to policy action probabilities And biased between the output layer and the pre-disaster action. Perform element-wise Hadamard product fusion Then, extract the actions (highest power omnidirectional transmission, protocol stack switching, medium power, lowest power silence).

[0050] Step S41, Counterfactual Differential Reward Calculation: Instead of a globally equal reward distribution, a differential reward mechanism is constructed:

[0051] In the formula, Represents a node At time step The reward value obtained for a single step of reinforcement learning; These represent the constant adjustment weight coefficients for each positive incentive and negative penalty; This is an indicator function; it is 1 when the data packet is successfully delivered, and 0 otherwise. This represents the average queuing congestion time reference value within the local one-hop range of the node; This represents the current actual queuing delay value; Represents a node At time step The specific communication actions taken; A function representing the physical energy cost of performing a specific action; This is an indicator function; it is 1 when a protocol stack switch triggers a hardware lock, and 0 otherwise. The reward mechanism stipulates that additional positive incentives are only given when the agent's actions effectively reduce the local average (i.e., improve congestion).

[0052] Step S42, the cross-temporal "cloud-based offline pre-training - post-disaster pure distributed fine-tuning" architecture: The cloud-based offline pre-training corresponding to step S12 in the pre-disaster phase is followed by purely distributed online fine-tuning in the post-disaster phase. The strategy network of this invention does not rely on a centralized training server after a disaster.

[0053] After the disaster, the system completely severed its connection to the Critic network and global state. The nodes retain only a very lightweight Actor network and rely entirely on "counterfactual differential rewards" for local fine-tuning to achieve policy emergence that adapts to the current real-world terrain.

[0054] This invention is based on the counterfactual differential reward mechanism and the dynamic evolution of the strategy, realizing "implicit spatial reuse" driven by counterfactual rewards. In high-density resettlement sites in disaster areas, nodes often transmit at full power to ensure their own connectivity, which in turn causes a serious broadcast storm at the MAC layer.

[0055] (1) Counterfactual differential reward: The evolution of policy networks abandons the global equal reward and instead uses "local one-hop average congestion delay" as the dynamic benchmark.

[0056] (2) Spontaneous cooperative emergence: Positive incentives can only be obtained when the actions of the agents (such as actively and silently relinquishing the channel) actually reduce the local mean and improve congestion. This prompts the network to spontaneously emerge cooperative strategies similar to frequency division / space division multiplexing through distributed game theory in the absence of a central scheduler to allocate frequencies.

[0057] (3) Purely distributed fine-tuning: The post-disaster system completely cuts off its dependence on the global state. The nodes retain only a lightweight Actor network and rely entirely on the aforementioned counterfactual local rewards to fine-tune locally, thereby achieving strategy evolution that adapts to the current real terrain.

[0058] In summary, this invention breaks away from the traditional purely reactive post-disaster framework. It solves the cold start problem by "pre-setting computing power before disaster + post-disaster probe filtering", overcomes the failure of high-density network perception by "introducing GNN with physical edge conditions", and finally guides nodes to spontaneously restrain selfish behavior through "counterfactual differential reward". Thus, it maximizes the lifespan and data delivery rate of emergency communication networks under extremely limited hardware and electromagnetic environments.

[0059] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0060] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0061] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause an electronic device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0062] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An emergency networking method integrating pre-disaster prior knowledge and edge condition graph neural networks, characterized in that, include: Step S1, Pre-disaster preparation: Quantification of pre-disaster status and spatiotemporal redistribution of computing power; Step S10, Quantification of multidimensional physical indicators: During daily operation, the terminal continuously collects six core physical indicators: absolute battery capacity, computing power, transmission power, connection stability, throughput, and packet loss rate. Multidimensional physical index extraction: During daily operation, continuously collect core index vectors from the terminal. Corresponding node Absolute battery capacity Computational ability Transmission power Connection stability Throughput and packet loss rate And perform extreme value normalization processing and map to interval; Step S11, Prior weight solidification: Calculate the prior score of the terminal's comprehensive performance using the Analytic Hierarchy Process (AHP), and generate an initial feature interaction prior weight matrix accordingly. Solidify the communication potential into a prior tensor and distribute computing power to edge nodes in advance. Step S12, offline pre-training in the cloud; In the pre-disaster phase, intensive pre-training is conducted. Relying on cloud computing power, a massive number of simulated disaster scenarios are constructed using domain randomization technology. The Actor network and Critic network complete global convergence of centralized training and distributed execution in the cloud, resulting in a mature basic policy model, which is pre-installed on all mobile terminals. Step S2: Post-disaster instantaneous survival verification and prior activation; In the moments after the disaster: heartbeat probes and rapid topology trimming; The network employs a negative migration prevention mechanism the instant it switches from normal operation to post-disaster operation, namely: Minimalist heartbeat verification: Triggers a millisecond-level survival probe with extremely low overhead the instant a disaster occurs; Each node will broadcast a minimalist heartbeat packet; Avoid convergence oscillations: If a node does not receive a heartbeat response from a high-potential node before the disaster, its prior confidence is instantly reduced to zero at the physical level, and the damaged node is quickly removed to prevent convergence oscillations in the early stages of the network. Step S3, Post-disaster perception: High-dimensional topological state perception based on edge conditional graph neural network EC-GNN; The physical reconstruction edge-conditional graph neural network EC-GNN uses the instantaneous signal-to-noise ratio and end-to-end queuing delay of the physical layer as edge conditions, and utilizes a multilayer perceptron (MLP) for feature concatenation and modulation; realistically reproducing the erosion of information entropy by electromagnetic fading; Extremely low-overhead quantization interaction: To prevent high-dimensional feature interactions from exhausting the bandwidth of the disaster area, message broadcasting is designed as an event-triggered mechanism, and high-dimensional floating-point numbers are compressed into low bits for discrete representation before being sent. Spatiotemporal smooth transition: Construct a control factor that decays exponentially with time step, and dynamically fuse the topological prior extracted before the disaster with the real-time channel state after the disaster to achieve a smooth transition from prior-dominated to real-time dynamic-dominated. Step S4, Post-disaster decision-making: Dynamic evolution of strategies based on counterfactual differential reward mechanism; Step S40, Action Mapping: node Using Actor Networks Mapped to policy action probabilities ， And offset the output layer from the pre-disaster action. Perform element-wise Hadamard product fusion ， The action is then extracted, which includes one of the following: highest power omnidirectional transmission, protocol stack switching, medium power, or lowest power silence. Step S41, Counterfactual Differential Reward Calculation: Instead of a globally distributed reward system, a counterfactual differential reward mechanism is constructed: In the formula ， Represents a node At time step The reward value obtained in a single step of reinforcement learning ； These represent the constant adjustment weight coefficients for each positive incentive and negative penalty; This is an indicator function; it is 1 when the data packet is successfully delivered, and 0 otherwise. This represents the average queuing congestion time reference value within the local one-hop range of the node; This represents the current actual queuing delay value; Represents a node At time step The specific communication actions taken; A function representing the physical energy cost of performing a specific action; This is an indicator function; it is set to 1 when a hardware lock is triggered by a protocol stack switch, and 0 otherwise. The counterfactual provision states: only when the node Additional positive incentives are only given when the action is significantly lower than the local average. Step S42, Cross-temporal cloud-based offline pre-training - post-disaster purely distributed fine-tuning architecture: The cloud-based offline pre-training corresponding to step S12 in the pre-disaster phase is followed by purely distributed online fine-tuning in the post-disaster phase. After the disaster, completely sever all access to the Critic network and global state. The nodes retain only the Actor network and rely entirely on the counterfactual differential reward mechanism to fine-tune locally, achieving policy emergence that adapts to the current real terrain.

2. The emergency networking method for integrating pre-disaster prior knowledge and edge condition graph neural networks according to claim 1, characterized in that, The specific process in step S11 is as follows: AHP Consistency Assessment and Generation of Prior Scores for Overall Performance: Construct a positive and negative judgment matrix By solving for the satisfaction The largest eigenvalue and its corresponding weight vector, satisfying the consistency ratio Under the premise of [the above], the prior score of the overall performance of the computing terminal In the formula Represents a positive-negative judgment matrix; This represents the weight vector corresponding to the indicator. This represents the largest eigenvalue of the judgment matrix; Represents the calculated nodes The overall performance prior score; Topological prior weights are fixed: Based on the prior scores of the node's overall performance, the initial feature interaction weight matrices used to guide the perception layer are calculated respectively. And the action bias weight vector used to guide the decision-making level For any pair of nodes (i,j) in the network topology that may generate data interaction, their initial feature interaction weights are... Defined as: In the formula, Represents a node For nodes Initial feature interaction weights; and This is a constant adjustment coefficient; and They are nodes and nodes The overall performance prior score; and They are nodes and nodes The multidimensional physical feature vector; This refers to the regional correlation coefficient. Represents a node The set of single-hop direct neighbor nodes; The index variable is used to traverse the neighbor set; the initial feature interaction weight function is not only proportional to the product of the prior scores of the two end nodes, but also integrates the cosine similarity of multidimensional spatial features and the regional correlation coefficient. By solidifying the communication potential into a prior tensor, the computing power is distributed to the edge nodes on the spatial axis, thus solving the cold start problem of the model after the disaster.

3. The emergency networking method for integrating pre-disaster prior knowledge and edge condition graph neural networks according to claim 1, characterized in that, In step S3, specifically: Post-disaster, nodes Extracting global high-dimensional space belief states using EC-GNN under limited observation conditions; Step S30: Construction of local observation vectors; node Sniffing to obtain observation vectors ; Self-characteristics Includes normalized power, protocol status, and node degree; Edge features Includes instantaneous signal-to-noise ratio of physical links End-to-end queuing delay ; Step S31, based on event triggering and low-bit quantization of side conditional message interaction: To avoid exhausting the precious physical channel bandwidth of the disaster area due to the interaction of high-dimensional feature vectors in GNNs, a communication overhead penalty and dimensionality reduction mechanism are introduced, specifically: In terms of feature dimensionality reduction, a multilayer perceptron is constructed to map the concatenated neighbor node state vectors and link feature vectors into a new message carrier to achieve dimensionality reduction. Regarding communication overhead penalties, a penalty term for performing high-energy-consuming actions or locking heterogeneous protocol stacks is explicitly introduced into the counterfactual differential reward mechanism. In the formula, This represents the action penalty weighting coefficient, used to adjust the penalty intensity for actions such as invalid broadcasts; Represents a node exist Perform actions at all times The resulting motion overhead; Represents a node exist The strategic actions taken at each moment are aimed at reinforcing the learning objectives to curb ineffective feature broadcasting; The event triggering is as follows: a node only triggers physical layer broadcast interaction of state information when there is a substantial change in the local network topology, including but not limited to when the link signal-to-noise ratio fluctuates drastically beyond a threshold or when the one-hop degree changes abruptly. The aforementioned low-bit quantization: Before broadcasting, the transmitting end uses a local quantizer to convert the high-dimensional floating-point features... The data is compressed into a low-bit discrete representation, which is then dequantized at the receiving end and subsequently used to generate a side-conditional modulation message through a multilayer perceptron; the low bit is either 4-bit or 8-bit. In the formula, This represents the aggregation layer index of the edge-conditional graph neural network EC-GNN; Indicates the first The edge-conditional modulated message feature vector generated by the layer; This represents the mapping function used to process message vectors in a multilayer perceptron network. This represents the dequantization operation function; Representing neighboring nodes In the The high-dimensional hidden state belief vector output by the layer; Step S32, Dynamic attention fusion and prior confidence time decay: To address the negative migration caused by backbone node damage, a time-stepping mechanism is introduced. Control factor of exponential decay ;Probe liveness indicator function Control factors The real-time attention score is weighted and fused to generate a dual-mode dynamic attention coefficient: In the formula, Indicates the first Layer nodes For neighboring nodes The dual-mode dynamic attention coefficient; and These represent the unnormalized real-time attention scores calculated from the local channel state; This represents the initial maximum control factor constant; This represents the decay coefficient, indicating the rate at which the control factor decreases with time step. Represents a node The set of single-hop direct neighbor nodes; To traverse the set of neighbor nodes index variables; and This is the heartbeat probe survival indicator function; its value is 1 if the node is alive, and 0 otherwise. The latter part of the formula introduces the survival indicator function to... By performing masking operations and renormalization, the prior weights are accurately redistributed proportionally among the surviving nodes. Step S33, self-embedding reinforcement and residual normalization; node Establish an independent, fully connected channel to enhance the weight of its own remaining battery power perception; Specifically: Nodes First, construct a system containing normalized remaining battery power. Current status of heterogeneous communication protocols and local node degree features Its own feature vector: This inherent feature is utilized by a dedicated weight matrix when passing through the network. Independent linear mappings are performed, residual connections are introduced, and LayerNorm layer normalization operations are introduced to stabilize gradient backpropagation in deep networks. The specific iterative update rule of the layer is described as follows: In the formula, This represents the mapped output of an independent, fully connected channel; after going through The iterative aggregation of layered EC-GNNs rigorously up-maps low-dimensional local observations that were originally limited by single hops to include global dimensions. High-dimensional belief state vector influenced by topological jumps .

4. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method described in claims 1 to 3.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method described in claims 1 to 3.

Images